A retailer trains a demand forecasting model on two years of weekly sales. After a new loyalty program launches, customers buy larger baskets at different points in the month. The model, code and features are unchanged, and the forecasts miss by a little more each week.
Distribution shift is a difference between the data a machine learning model was trained on and the data it receives in use, which can lower performance even when the model itself has not changed.
The authors of the WILDS benchmark paper, Koh et al., describe distribution shifts as cases where the training distribution differs from the test distribution, and note that such shifts can substantially degrade the accuracy of machine learning systems deployed in the wild. They also observe that these shifts are under-represented in the datasets widely used in research.
Researchers name types of shift by what changes. In Detecting and Correcting for Label Shift with Black Box Predictors, Lipton, Wang and Smola study label shift, where the frequency of each outcome changes while the way each outcome appears in the inputs stays the same.
How is distribution shift different from data drift and concept drift?
- Distribution shift is the general term for any mismatch between the data a model was trained and evaluated on and the data it sees in use.
- Covariate shift means the inputs change while the relationship between inputs and outcome stays the same. Label shift means the outcome frequencies change.
- Concept drift means the relationship itself changes, so the same input should now lead to a different prediction.
- Data drift is a monitoring term for changes in input data observed over time in production.
Where does distribution shift break in AI work?
Shift rarely produces an error message. The pipeline runs, predictions arrive and quality falls without a clear signal.
- Stale evaluation: a model keeps passing a test set built from last year’s data while live performance drops.
- New segments: a product launch, a new region or a policy change brings inputs the training data barely covered.
- Retraining without a record: a monthly retraining job overwrites last month’s prepared input, so when a metric drops the team cannot compare the two states.
CUBIG Shorts
No Error. The Metric Still Dropped.
Episode 7 of It Worked in the Demo, CUBIG’s webinar series on enterprise AI.
How do teams detect and respond to distribution shift?
Teams compare feature and label distributions between the training data and recent data with statistical tests, and track model performance on recently labeled examples where labels are available.
When shift appears, the first question is what changed: the data, the labels, the model configuration or the evaluation set. Keeping versioned datasets and fixed evaluation conditions lets a team compare results before and after the change, then decide which prepared data still applies and what needs to be prepared and validated again. Common responses include retraining on recent data, reweighting samples and adding data that covers the new conditions.
Related entries: model drift, reproducibility and schema drift.